Online Reviews in SKU-level Demand Forecasting Using Global Cross-Learning LSTM

Abstract

Background: Daily SKU-level demand forecasting is challenging in zero-inflated e-commerce data because absolute-error optimization can favor zero forecasts and fail to capture demand activation. Objective: This study evaluates a Global Cross-Learning LSTM that combines transaction history, structural sparsity markers, and online-review signals as behavioral proxies. Methods: Computational forecasting experiments used daily SKU-level transaction and review data from the Amazon and Olist datasets. Four ablation configurations (M0–M3) were compared with SES, Croston, and SBA using MASE, RMSSE, and AMSE. The LSTM used a 30-day lookback window, 50 units, the Adam optimizer, a batch size of 128, 20 epochs, and Huber loss. Paired differences across 22 entities were assessed using the Wilcoxon signed-rank test. Results: Across the full timeline, M0 obtained the lowest MASE (0.4243), reflecting the advantage of zero forecasts on inactive days. During activation windows, however, M3 reduced AMSE from 0.7769 to 0.6195 and produced a statistically significant improvement in RMSSE (p < 0.05). Global pooling also supported forecasting for cold-start and lumpy-demand items. Conclusion: The findings support Proxy Theory by indicating that review volume and valence can provide leading information before sufficient transaction data accumulate. Theoretically, the study links behavioral proxies with global representation learning for intermittent demand. Practically, the model can inform replenishment and safety-stock decisions by reducing underforecasting during demand activation. The novelty lies in jointly evaluating online-review proxies and global cross-learning at the SKU level under lifecycle-specific conditions.